Electrical load forecasting models: A critical systematic review. (November 2017)

• Record Type: Journal Article
• Authors: Kuster, Corentin; Rezgui, Yacine; Mourshed, Monjur
• Is Part Of: Sustainable cities and society. Volume 35(2018)
• Page Start: 257
• Page End: 270
• Publication Date: 2017-11
• Subjects: Electric consumption and load prediction -- Forecasting models -- Machine learning -- Regression -- Time series analysis -- Long-term/short-term forecasting; Sustainable urban development -- Periodicals; Sustainable buildings -- Periodicals; Urban ecology (Sociology) -- Periodicals; 307.76
• DOI: 10.1016/j.scs.2017.08.009
• Languages: English
• ISSNs: 2210-6707

Abstract

Highlights: A systematic review protocol provides unbiased and meaningful meta-information. A direct model accuracy comparison across studies is meaningless. A taxonomy for an informed forecasting model's selection is proposed. Recommendations on writing electrical load forecasting related paper are given. Abstract: Electricity forecasting is an essential component of smart grid, which has attracted increasing academic interest. Forecasting enables informed and efficient responses for electricity demand. However, various forecasting models exist making it difficult for inexperienced researchers to make an informed model selection. This paper presents a systematic review of forecasting models with the main purpose of identifying which model is best suited for a particular case or scenario. Over 113 different case studies reported across 41 academic papers have been used for the comparison. The timeframe, inputs, outputs, scale, data sample size, error type and value have been taken into account as criteria for the comparison. The review reveals that despite the relative simplicity of all reviewed models, the regression and/or multiple regression are still widely used and efficient for long and very long-term prediction. For short and very short-term prediction, machine-learning algorithms such as artificial neural networks, support vector machines, and time series analysis (including Autoregressive Integrated Moving Average (ARIMA) and the Autoregressive Moving Average (ARMA)) are favoured. The most widely employed independent variables are the building and occupancy characteristics and environmental data, especially in the machine learning models. In many cases, time series analysis and regressions rely on electricity historical data only, without the introduction of exogenous variables. Overall, if the singularity of the different cases made the comparison difficult, some trends are clearly identifiable. Considering the large amount of use cases studied, the meta-analysis of the references led to the identification of best practices within the expert community in relation to forecasting use for electricity consumption and power load prediction. Therefore, from the findings of the meta-analysis, a taxonomy has been defined in order to help researchers make an informed decision and choose the right model for their problem (long or short term, low or high resolution, building to country level).